An AI assistant reading API documentation through a secure gateway

MCP server for API docs: give AI assistants access

AI coding assistants are only as good as what they can read. Ask one to call your internal billing API and it will guess field names. SelfDocHub has a built-in MCP server (Model Context Protocol), so an assistant reads the real, current documentation instead: the endpoints, the schemas, the version history and the notes your team wrote.

What the assistant can do#

Through MCP an assistant can list services and endpoints, read the parameters and responses of an endpoint with the schemas they use, search the documentation including your text pages, browse the version history, read the change log and diff two versions, and read your Markdown overlays. The tools include list_services, list_endpoints, get_endpoint_schema, search_docs, list_versions, get_version_changelog, diff_versions and get_overlay.

By default the integration is read-only. It never exposes license data and never creates, edits or deletes anything.

Scoped tokens#

Access is granted by tokens, and any number can be active at once. Each token has its own settings, so one client can see a couple of services while another sees more:

  • the services it may read: any other service is refused by every tool;
  • an IP/CIDR allowlist;
  • the maximum number of search results;
  • whether overlay text is included in answers;
  • independent switches for the schema, version-history and search tool categories;
  • whether the model may see other branches and historical versions.

The secret is generated by the server, shown once and stored only as a salted hash. Replacing it invalidates the old one at once, and changes to a token take effect immediately without a restart. The page also shows when each token was last used.

Connect a client#

In an MCP-capable client, such as a connector setting in Claude or Cursor, enter the URL https://<your-domain>/mcp and the header Authorization: Bearer <token>. A quick check with curl works too:

curl -X POST https://<your-domain>/mcp \
  -H "Authorization: Bearer <token>" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

Optional: Try-It through MCP#

If you want the assistant to verify an answer with a real call, you can enable Try-It for a token and open selected environments to it. This is a potentially dangerous feature, so it is off by default and has its own controls: the model can only use environments an administrator opened to it, with the credentials configured there, and limits keep it in check. Try it on test environments first, keep the limits low and prefer tokens with a narrow list of services. Try-It through MCP is part of the Business and Unlimited plans.

Why self-hosted matters here#

Your API documentation describes your internals. With SelfDocHub the server that answers the assistant is your own, behind your own network rules, with an optional IP allowlist per token, and every rejected request is logged. The request quota depends on the plan: 20,000 requests per month on Free, 100,000 on Team and 500,000 on Business.

Frequently asked questions

Can the AI change or delete anything?

No. By default the integration is read-only. It never exposes license data and never creates, edits or deletes anything.

Can I limit what an AI client sees?

Yes. Each token has its own list of allowed services, an optional IP allowlist, a search result limit and switches for the schema, version-history and search tools.

Can the model send real requests to my API?

Only if you enable the separate Try-It mode for a token and open specific environments to it. It is off by default and is a licensed feature of the Business and Unlimited plans.

Run it on your own infrastructure

The Free plan includes a license for one installation. Start with Docker in a few minutes, no credit card needed.